In a recent White Paper on Drug Development, the FDA has identified computer modeling as one of the new enabling technologies for improving drug development processes. Computational neuropharmacology, as described herein, is different from database management, data mining and pattern recognition, in that it attempts to predict outcomes based on mathematical models starting from well-defined physico-chemical principles, rather than applying statistical inference techniques.
Database mining is very useful, especially when applied to the integration of pharmacological with clinical information, but by itself it is usually not sufficient to identify drug targets with a sufficient probability of affecting a disease.
As an example, correlating binding affinity data on various human receptor subtypes for a number of neuroleptics with their clinical side-effects has enabled the identification of a number of key receptors involved in weight gain [56]. However, this particular approach uses only binding affinity data and does not take into account the actual dosage used in the clinical setting. Also, the interaction of neuroleptics at the receptor subtypes is also modulated by the affinity of the endogeneous neurotransmitter for the same receptor subtype, the presynaptic firing pattern, possible negative feedback via presynapic autoreceptors and the presence of pharmacologically active metabolites. Thus a mathematical approach encompassing all these interactions in a quantitative way is clearly necessary and has not been presented.
Another major issue is estimating the functional concentration of antipsychotic drugs in the brain. Fortunately, PET imaging using specific radioactive tracers can be used to determine the competition with added antipsychotic drugs at specified receptor subtupes, such as the dopamine D2R. A number of studies have made possible the prediction of brain D2R occupancy levels, measured by radio-active tracer displacement, in function of the plasma level [97], results that have been actually confirmed by experimentally in the human brain [94]. However, this approach as presented does not enable one to determine the actual concentration of the neuroleptic and its metabolite in nM in the human brain, which can be used at other synapses beyond dopaminergic synapses. In addition, it is not possible to determine the level of postynaptic activity, especially in the case of partial agonists.
Simulation has been applied to the problem of slow antagonist dissociation and long-lasting in vivo receptor protection; see Vauquelin et al. [99] which addresses the time-dependent evolution of receptor occupancy, dependent upon both affinity and half-life of the drug. Such approaches while interesting, do not address the level of receptor activation or determination of actual brain drug concentration in clinically relevant conditions.
At least one attempt to simulate the outcome of clinical trials with the antipsychotic quetiapine has been published; see Kimko et al. [51]. In this study, the relation between plasma drug concentration and BPRS scale as a measure of clinical efficacy was modeled as a linear, U-shaped, inhibitory Emax and sigmoidal Emax function, using standard statistical analyses. There was no attempt to include actual physiological interactions or pathology parameters. In view of the large discrepancy between drug plasma levels and actually measured functional brain receptor occupancies [94], it was of no surprise that the model deviated quite considerably from the actual data.
Another commonly used rule for determining the clinical efficacy or side-effect liability of neuroleptics is the degree of D2R occupancy, measured with tracer displacement. This is given byD2R−occ=D2R−occmax*Dose/(Dose+Ki)                Where Dose is the dose of the neuroleptic and Ki is a parameter determined from PET imaging studies with radio-active tracers, usually 11C-racopride.        This method at best yields modest correlatiuons and is clearly not adapted to predict the clinical outcome of novel therapeutic agents with partial agonist effects at the D2R, such as aripiprazole.        
The computational neurosciences approach is based upon mathematical models describing actual biophysical processes. In most cases, the readout of these models is neuronal action potentials, which obviously is related to behavior. This area of research has yielded an enormous array of novel insights in the way neuronal circuits code information and perform certain cognitive operations. However, prior to this invention these models have not been integrated together in practical way which would represent a disease state and predict the clinical effect of drugs thereon.
With regard to the problem of schizophrenia, the major brain regions involved are the cortex, the striatum and the hippocampus. There is evidence of a dysfunctional signal-to-noise ratio observed with EEG techniques in patients and siblings [105]. According to the hypothesis developed by Grace [39], the ventral striatum integrates inputs from cortical, amygdale and hippocampal regions. In order to develop an adequate mathematical model of the deficient information processing in the pathology of schizophrenia, there is a need for a detailed model of each of these brain regions, followed by an integration of the different inputs into the ventral striatal computation unit.
There are computational models in the neural areas of interest that have a basic science focus that are linked to schizophrenia. When human outcomes are of interest, the results are more generally linked to behavior rather than disease or particularly schizophrenia (e.g., Montague et al. [70] and Smith et al. [86]). For example, there have been computational models of the striatum and medium spiny projection neurons (e.g. Wolf et al., 2005), of dopamine signaling (e.g., Schultz et al. [83]), of cortical circuitry (e.g., Chen [21]), and of information processing in the frontal cortex and basal ganglia (e.g., Amos, 2000). None of those computational approaches, however, integrate a biologically-based simulation to human clinical data and derive therapeutic drug targets. The previous modeling did not correlate and compare the output of the simulations with clinical trial results to predict clinical outcomes for untested drugs or determine targets for novel therapeutic drugs.
A range of models have been developed directed at cortical and hippocampal processes relevant to schizophrenia (for review, 32). Recent modeling efforts have largely focused on understanding three fundamental cognitive processes: working memory, the decision process, and attention.
Based on the pioneering work of Pat Goldman-Rakic [37] and others, much research has focused on deficits in working memory function in prefrontal cortex (PFC) and its contribution to loss of executive control and disorganized thinking in schizophrenia. The major modeling efforts in this direction attempt to account for the ability of cortical networks to maintain a stable pattern of activity—e.g., a stable firing pattern—across a population of neurons, in the face of distracting stimuli. Durstewitz and colleagues [30] showed that a simple cortical architecture, based on interconnected pyramidal cells and interneurons, could maintain stable firing (so-called attractor behavior) with firing rates corresponding to experimental data from PFC, and that dopamine, acting through D1 receptors worked to stabilize the activity when distracting activity patterns were applied to “knock” the “goal” out of working memory. A succession of models has taken this attractor model further, notably X J Wang and Miller et al. [100, 67] have incorporated more accurate neuronal models with additional biophysical properties, and have shown more sophisticated attractor behavior. Additional interneuronal types have also been incorporated into the cortical network, based upon a different functional role for dendrite-targeting calbindin positive interneurons versus soma targeting palvalbumin positive interneurons, versus the calretinin positive interneurons that target other interneurons (100).
Thus, a substantial body of computational modeling has been developed to account for recent experimental findings on working memory, decision making, attention and other cognitive processes. The biophysical models of neurons used in these models vary from extremely simplified (integrate-and-fire models), to somewhat sophisticated multicompartmental models incorporating 6 or 7 intrinisic currents and several synaptic receptor currents (usually AMPA, NMDA, and GABAA).
A final set of relevant work in the literature deals with the action of neuromodulators on brain circuit function. In a number of models 14, 27, 92, 31) however, the effect of dopamine is not explicitly calculated on model parameters it is just assumed, loosely based on experimental literature, that dopamine increases NMDA conductance by a fixed percentage. Similarly, in Hasselmo's models of the effect of acetylcholine on functional connectivity in various hippocampal pathways [44], the effect of the neuromodulator is introduced only at a single effective concentration, and only through its effect on a model parameter, such as synaptic weight.
The dynamics of the neuronal circuits in the prefrontal cortex are important for addressing the issue of cognitive deficits in schizophrenia, an area which has recently been recognized by the NIMH and the FDA as a major unmet medical need. In fact, a very low fraction of ‘stable’ schizophrenia patients are able to return to their level of professional activity.
It is clear that for accurately describing the effects of antipsychotics and therapies used in psychiatric diseases on the neuronal dynamics in the prefrontal cortex, a new model is required integrating the pharmacological effects of other neurotransmitter systems, such as acetylcholine, serotonin and norepinephrine are needed.
With the advent of pharmacogenomics and functional data on patient genotypes, it is mandatory to have a model which can incorporate these functional genotypes in a rational way.
In the striatum, the most important cell type is a medium spiny neuron (MSN) GABA cell, a computational model of which has been published (40). This model has a dopaminergic D1 mediated input and a glutamate afferent input and was intended to demonstrate a dopamine-induced bifurcation and the link to expected reward. However, as all neuroleptics effective in schizophrenia antagonize the dopamine D2R, this model—which lacks D2 receptor—is clearly not sufficient to describe the effect of dopamine receptor modulation on the relation between incoming glutamatergic signals and outgoing GABAergic signals. In addition, both 5-HT2C [28] and D(3) receptor ligands (110) have been documented to modulate the dynamics of dopamine release in the striatum. At least these two pharmacological influences need to be incorporated in the model, as many neuroleptics also affect these receptor subtypes. In addition, novel insights in the pathophysiology of schizophrenia point to the idea of signal and noise (105). Also the gating of hippocampus and amygdala which is seen as major inputs into the N. accumbens [39] is not implemented in this model. Recent experimental data indeed have provided information on the electrophysiological interaction between prefrontal cortex, hippocampus and N. accumbens [38].
There is a need for the integration of information on the interaction and pathways of the different neurotransmitter circuits with their different receptor subtypes in the human brain. Information is available from preclinical microdialysis and voltammetry studies on neurotransmitter levels and electrophysiological studies in well-defined brain regions in the monkey and rat brain. Human information relates imaging data (PET, MRI), functional genomic and postmortem data in appropriate patient populations. The specific affinity of each drug for different receptor subtypes defines its interaction with various neurotransmitter systems.
In particular, interactions between serotonin and norepinephrine are very important in the setting of schizophrenia. Indeed, many antipsychotics currently in use have pharmacological effects at serotonerge and noradrenerge receptor subtypes. In addition, a number of drugs used to treat depression have serotonergic and/or noradrenergic modulation as their primary mode of action. However, good models which can accurately describe the interaction between these two types of neurotransmitters are lacking, despite a wealth of information on the reciprocal interaction between Locus Coeuruleus, the source of noradrenergic neurons and the Dorsal Raphe Nucleus, the source of serotonergic neurons. Therefore a good mathematical model which can take all these interactions into account is necessary.
Computational neuroscience models have described the calculation of biomarkers such as fMRI [60] and EEG [64]. However, other parameters of interest in the field of drug development have not yet been published. There is a need for ways to adapt computational neuropharmacology approaches to (1) identify the ‘ideal profile’ of drugs, (2) estimate the effect of comedications, (3) perform power calculations based not only on pharmacokinetic variability, but on pharmacodynamic variability as well, (4) estimate the clinical effects of specific functional genotypes, and (5) estimate the influence of chronopharmacodynamics (i.e. the time of the day when drugs are given).